PageSpeed Insights scores rose on pages nobody touched
Summary
PageSpeed Insights now runs Lighthouse 13.6.0, whose updated load simulator pushes lab scores up without anything on the page changing.
Before reporting any score gain, check that both reports name the same Lighthouse version.
PageSpeed Insights now runs Lighthouse 13.6.0, which changes how the tool simulates a page load rather than anything about how pages are built. Barry Pollard, Chrome’s web performance developer advocate, said on Bluesky that the new version produces more accurate simulated First Contentful Paint and Largest Contentful Paint and will generally give higher scores, per Search Engine Journal’s report. A page can therefore score better than it did last month with nothing on it changed.
Lantern is the simulator behind Lighthouse’s estimates: Lighthouse loads the page at full speed, then works out how it would have loaded on a slower connection. The 13.6.0 release notes, published on 27 October, list several changes to Lantern, and two of them bear directly on FCP and LCP. Lighthouse 13.5 used each request’s fetch priority to decide whether that request blocked rendering. Lantern now reads Chrome’s own render-blocking label for each request and falls back to priority only when the label is missing. The second change narrows what feeds the LCP estimate. When the LCP element is text that appears in the same frame as the page’s first content, Lantern now estimates LCP from the same requests and page tasks it uses for FCP, instead of from a set that can include every request made before LCP.
Ivan Akulov, a web performance consultant whose GitHub issue kicked off some of the work, showed what the old estimate did to a test page built with Framer whose only above-the-fold content was text. A single element counted as both FCP and LCP there, and Lighthouse still reported LCP much later than FCP. Preloading the page’s JavaScript modules made the real page interactive sooner and made the simulated LCP worse. On that page the simulator was marking down a change that helped users, and the Lighthouse team closed Akulov’s issue as completed with the 13.6 upgrade.
Pollard said the change is not specific to single-page apps, though JavaScript-heavy sites are more likely to feel it. He was quoting a developer who had posted that SPAs, react.dev among them, improved sharply after the update. The safer assumption going into an audit is that a build shipping a lot of JavaScript moved more than a mostly static page did, and that the tool explains the movement until the field data says otherwise.
Any two PSI runs that straddle the switch measured the page with different simulators. A monthly performance report whose previous run predates 13.6 is comparing two methods, and a chart tracking lab scores over time will show a step at the version change that no deploy caused. Every PSI report names the Lighthouse version it used, next to the emulated device, which is the only way to tell which series a number belongs to.
Field data in the same report is untouched by all of this, because it comes from real Chrome visits over the previous 28 days through the Chrome User Experience Report. Google’s own PageSpeed Insights documentation says good lab data does not necessarily mean real-user experiences will also be good. For single-page apps, real-user coverage is also improving separately through Chrome’s experimental CrUX soft navigation fields.
What to do
Re-baseline before the next report goes out. Re-run the pages you track, treat the pre-13.6 numbers as a closed series rather than smoothing the step out of the chart, and note the version alongside each score from now on. Any claim that users got a faster page should come from the 28-day CrUX figures: if field LCP did not move, the score gain was a measurement change. Flag the jump in the next client report yourself, before someone reads it as the result of work you did not do.
Watch out for
- DevTools and PSI disagreeing. Lighthouse expects 13.6 to ship in Chrome 159’s DevTools, which Chrome’s release schedule lists for stable release on 8 December, so a stable DevTools run and a PSI run on the same page can return different numbers until then.
- More movement to come. An open Lighthouse issue is looking into cases where Lantern’s simulated LCP runs far above measured LCP, with large JavaScript and SPA sites as the examples, so JavaScript-heavy pages are not necessarily finished moving.
- A changed verdict on agentic discovery. The same release updated the Agentic Resource Discovery audit, which looks for a site’s ARD manifest, to check for an Agentmap line in robots.txt, a rel=“ard” link tag or Link header, and /.well-known/ard.json before the older ai-catalog names. A site that had already moved to ard.json and showed Not Applicable under 13.5 now gets a real result.